Personalisasi Cold-Start pada Sistem Pembelajaran Adaptif tanpa Embedding Berbasis Pemodelan Penguasaan Konsep dan Multi-Armed Bandit

Syafei, Farros Hilmi (2026) Personalisasi Cold-Start pada Sistem Pembelajaran Adaptif tanpa Embedding Berbasis Pemodelan Penguasaan Konsep dan Multi-Armed Bandit. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025201012-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (8MB) | Request a copy

Abstract

Indonesia menetapkan Koding dan Kecerdasan Artifisial sebagai mata pelajaran mulai tahun ajaran 2025/2026, sementara hanya 5,31 persen sekolah dasar memiliki akses komputer untuk pembelajaran. Karena itu, ribuan sekolah memulai mata pelajaran yang sama tanpa satu pun riwayat interaksi pelajar, dengan anggaran komputasi terbatas, dan dengan guru yang perlu memeriksa alasan tiap rekomendasi sebelum meneruskannya. Model knowledge tracing mutakhir menuntut korpus interaksi berskala besar sebelum dapat dijalankan, dan meringkas pelajar menjadi embedding yang tidak dapat ditunjuk maupun diperiksa. Penelitian ini mengusulkan HCIE (Hierarchical Cognitive Inference Engine), sebuah Intelligent Tutoring System yang menyimpan penguasaan sebagai skor per konsep yang melekat pada nama konsepnya menggunakan filter Kalman skalar, memilih konsep dan bentuk penyajian berikutnya dengan Multi-Armed Bandit (Thompson Sampling), dan menyaring kesiapan melalui graf prasyarat kurikulum. Sistem berjalan sejak jawaban pertama, tanpa pelatihan, dengan biaya pembaruan konstan tiap interaksi. Pada kondisi tanpa riwayat, sistem menghasilkan estimasi yang informatif sejak jawaban pertama seorang pelajar, dengan AUC 0,62 sampai 0,76. Sementara itu, BKT tanpa pelatihan berada pada tebakan acak sebesar 0,500 dan model berbasis embedding belum dapat dijalankan sama sekali. Pada evaluasi gabungan dataset Junyi 2015 tanpa pelatihan luring, selisih ketelitiannya terhadap keempat model pembanding adalah +0,0088 (AUC 0,6051 terhadap BKT 0,5963, DKT 0,5892, SAKT 0,5730, dan GKT 0,5711). Karena itu ketelitiannya setara meskipun diperoleh tanpa satu pun tahap pelatihan. Karena penguasaan disimpan secara eksplisit pada tiap konsep, penelitian ini memperkenalkan ADC (Adaptive Dimension Controller), yaitu instrumen yang memeriksa apakah tiap faktor pedagogis benar-benar membawa sinyal pada data yang dihadapi sistem. Dari enam faktor, empat aktif dan dua dorman pada 96.727 interaksi. Pemeriksaan semacam ini tidak dapat diajukan kepada model berbasis embedding. Pada dimensi transfer, sebagian besar hubungan yang teramati dapat dijelaskan oleh faktor pengganggu. Karena itu yang dilaporkan adalah sisa sebesar 0,053 dengan p < 0,001, dan bukan hubungan sebab-akibat. Seluruh keputusan dicatat melalui event-sourcing sehingga dapat direkonstruksi dan diaudit ulang. Penelitian ini mengevaluasi mekanisme, dan tidak mengukur dampaknya terhadap capaian belajar.
=======================================================================================================================================
Adaptive learning systems commonly face the cold-start problem due to limited user interaction history. The condition is most acute for institutions newly adopting a curriculum or platform, when no historical data is available, computing budgets are limited, and system decisions are expected to be explainable to educators. This study hypothesizes that an explicit, auditable knowledge-state representation is sufficient to achieve competitive cold-start personalization without per-learner embeddings. To test this hypothesis, this study proposes HCIE (Hierarchical Cognitive Inference Engine), which represents concept mastery as an explicit low-dimensional knowledge state using a scalar Kalman filter and a Multi-Armed Bandit policy (Thompson Sampling) to deliver personalization from the very first interaction. In cold-start prediction experiments on the Junyi 2015 dataset without offline training, the approach attains an aggregate AUC of 0.6051, which is competitive with BKT (0.5963), DKT (0.5892), SAKT (0.5730), and GKT (0.5711). The margin is thin but remains positive under greater statistical power, namely an HCIE−BKT difference of +0.0088 at n = 10 and +0.0125 at n = 76 learners, so the approach is confirmed feasible and competitive. The accuracy comparison is positioned as a feasibility check, whereas the case for the approach rests on its practical properties, namely that it operates from the first interaction without historical data, that every decision is explainable and auditable, and that the update cost per interaction is constant. Because mastery is modelled explicitly per concept rather than as an embedding, this study is able to introduce ADC (Adaptive Dimension Controller), an observational instrument that tests whether the factors driving adaptive decisions are genuinely supported by empirical data. Such a test cannot be posed to an embedding-based model, so ADC also serves as evidence for the design choice itself. On the transfer dimension, ADC shows that most of the observed association can be explained by other factors. What is reported is therefore only the part that those factors cannot explain, namely 0.053 with p < 0.001, and not a cause-and-effect relationship. These findings support the research hypothesis. The entire approach is realized through an event-sourced engineering foundation that serves as an enabling technology supporting decision traceability, experimental reproducibility, and governance observability.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Adaptive learning systems, Cold-start personalization, Event-sourcing, Interpretable models, Knowledge tracing, Model yang dapat diinterpretasikan, Multi-armed bandit, Pendidikan pemrograman, Personalisasi cold-start, Programming education, Sistem pembelajaran adaptif
Subjects:	T Technology > T Technology (General) > T57.5 Data Processing T Technology > T Technology (General) > T57.8 Nonlinear programming. Support vector machine. Wavelets. Hidden Markov models. T Technology > T Technology (General) > T57.84 Heuristic algorithms. T Technology > T Technology (General) > T59.7 Human-machine systems.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Farros Hilmi Syafei
Date Deposited:	27 Jul 2026 03:43
Last Modified:	27 Jul 2026 03:43
URI:	http://repository.its.ac.id/id/eprint/138304

Actions (login required)

View Item